New Ignition Index metric measures global workspace dynamics in language models

Researchers have introduced the Ignition Index (I), a new metric designed to quantify global workspace dynamics within language models. This scalar metric operationalizes predictions from Global Workspace Theory (GWT) by analyzing per-layer accuracy transitions in response to input signal strength. The Ignition Index has demonstrated a 9.6-fold selectivity for genuine linguistic structure over spurious probe capacity across various transformer models and SSMs. Findings indicate that feedforward transformers generally exhibit higher ignition than SSMs, and recurrent architectures like Huginn-3.5B show distinct ignition patterns along their iteration axis compared to their depth axis. AI

IMPACT Introduces a novel metric for understanding internal dynamics of LLMs, potentially aiding interpretability research.

RANK_REASON The cluster contains an academic paper detailing a new metric for analyzing language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Global Workspace Theory
• Hugging Face
• Huginn-3.5B
• Ignition Index
• Mamba
• Pythia-410M
• transformer language models

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →